Summary
The article argues that Near Packaged Optics (NPO) is a practical intermediate step between traditional pluggable optics and full Co-Packaged Optics (CPO), driven by the need to shorten electrical paths in high-speed AI networking (especially scale-up domains like Rubin Ultra/NVL576). It implies that as NPO scales, significant dollar flow will go into optical engines, silicon photonics, lasers, connectors, and related passive photonics, but the article does not name specific publicly traded companies or disclose any positions.
-
•NPO moves the optical engine closer to the GPU/ASIC to reduce signal degradation, power draw, and board complexity at 200G+ per lane.
-
•Copper links become harder to push as scale-up networks grow and lane rates increase, creating the need for NPO.
-
•A 576-GPU system like Rubin Ultra could use 2.25 to 4.0 optical engines per GPU, implying 1,296 to 2,304 optical engines per system.
-
•NPO is described as a bridge to full CPO, offering more modular qualification, repair, and manufacturing flexibility.
-
•The next AI networking wave includes both scale-out (Ethernet fabrics) and increasingly scale-up (high-bandwidth, low-latency GPU-to-GPU links).
-
•Dollars in an NPO ramp flow into optical engines, silicon photonics, drivers, TIAs, photodetectors, lasers, VCSELs, external light sources, fiber attach, connectors, and passive photonics.